fix: normalize firmware reference value collection across baremetal/Azure

[butler54]

Summary

Fixes #113. Addresses the veritas/Rekor reliability issue reported in production, the Azure "external" reference values gap, and normalizes several scripts into the Makefile.

1. Multi-arch reference value collection (Fixes #113)

scripts/collect-firmware-refvals.sh's --tee now defaults to both instead of tdx: it runs veritas once per TEE and merges the results into a single output. TDX/SNP reference values use disjoint, TEE-prefixed keys (confirmed via trustee-chart's RVPS template, which already passes through whichever TEE-specific keys are present), so the merge is a safe JSON union. Applies to both the azure and baremetal platform branches — this also means make collect-firmware-refvals/make collect-azure-refvals now naturally produce complete reference values for heterogeneous/mixed-TEE deployments (e.g. values-baremetal-hub.yaml's kbs.tdx.enabled + kbs.snp.enabled) without any manual merge step. Explicit --tee tdx/--tee snp remains available for single-arch collection.

2. Rekor verification failures

Traced the actual failure to veritas's own container.py: Red Hat's cosign public key is hardcoded (not overridable via any flag), and --rekor-url only changes which Rekor server is queried for a transparency-log entry. Red Hat signs and logs these images against its own private Rekor instance, so pointing at a different one (e.g. public Sigstore) would not find that entry and fail — not a viable workaround.

The only flag that actually avoids this while still verifying the real signature is --skip-tlog, which is not present in the veritas version (v0.1.2) vendored by the currently-pinned quay.io/openshift_sandboxed_containers/coco-tools:0.5.1 container (confirmed via that image's requirements.txt; --skip-tlog was added to veritas after v0.1.2 was tagged, and no newer coco-tools tag exists yet).

Fix: switched to a host-installed veritas CLI (pip install "osc-veritas[snp]==0.1.3rc1" — published on PyPI, confirmed via a real install + veritas --help that --skip-tlog is present) instead of the container. Default --skip-tlog on the Azure branch; new --verify-tlog flag opts back into full verification. New prerequisite checks for veritas and cosign >= 2.0 (Azure only — bare metal verifies via oc adm release info --verify, not cosign/Rekor at all).

This host-install is a deliberate, temporary deviation — filed #132 to track moving back to the container once coco-tools ships a release with a newer veritas, with the reasoning for why container-based is preferred long-term (encapsulation, less host-environment drift, reproducibility).

Also deleted scripts/get-pcr.sh: the legacy manual predecessor to veritas-based collection, which depended on the same private Rekor instance but never actually completed verification (its cosign verify call was already dead/commented-out code) — strictly worse than the current path, and explicitly superseded per the docs' own description.

3. Azure "external" reference values

trustee-chart's kbs.azure.* block (snpLaunchMeasurement, smtEnabled, tsmeEnabled, abiMajor, abiMinor, singleSocket, smtAllowed) — static Azure-platform/VM-series constants, not something veritas collects — had no override wiring in coco-pattern at all, silently inheriting whatever trustee-chart's own defaults happened to be. Added overrides/values-trustee-azure.yaml (wired into values-azure.yaml) pinning these explicitly, plus documentation of what they are, the current VM-series mapping, and the caveat that the corresponding policy checks are currently commented out upstream (dormant, not enforced).

Filed #133 to track the actual collection/verification tooling (SSH into podvm + snpguest --openhcl) — scoped out since it needs live Azure hardware to validate, which wasn't available while authoring this PR.

4. Makefile/script normalization

Surveyed scripts/*.sh for consistency with the existing convention (thin wrapper-to-script for nontrivial scripts, vs. embedded shell directly in the Makefile for simple one-offs like detect-hardware/cache-keys):

• Inlined get-pccs-node.sh (10 lines, single oc query) into a new make get-pccs-node target; deleted the script.
• Added missing wrapper targets for scripts that had none: make gen-secrets, make check-pck-expiry, make check-collateral-expiry (scripts stay — too much python-heredoc/multi-step logic to reasonably inline).

Test plan

• Full scenario-test harness with stub veritas/cosign binaries: --tee both merge correctness (both azure and baremetal), single---tee opt-out, --hw-xfam-allow only added for the TDX run (not SNP), --skip-tlog on by default for azure only, --verify-tlog opt-out, invalid --tee/--platform rejected, missing-veritas/missing-or-old-cosign prerequisite errors, cosign not required for baremetal.
• bash -n on the modified script; make -n dry-run on all new/changed Makefile targets; ran make get-pccs-node for real (fails gracefully without cluster access, as expected).
• YAML validity of values-azure.yaml / overrides/values-trustee-azure.yaml.
• markdownlint / textlint clean on all touched docs (matching the CI-pinned tool versions from prior sessions).
• Live veritas/cosign invocations against real Azure/bare-metal clusters — not possible in this sandbox (no cluster/network image pull access). Will need validation against a real Azure cluster before merge.

Follow-ups filed

• chore: move collect-firmware-refvals.sh back to container-based veritas once coco-tools ships a newer release #132 — move back to container-based veritas once coco-tools ships a newer release (with rationale for why that's the preferred long-term state)
• feat: automate Azure SNP platform-constant collection via snpguest --openhcl #133 — automate Azure SNP platform-constant collection via snpguest --openhcl